Greetings, everyone! I’m covering two topics in today’s newsletter that many members have been asking about.

The first is the use of GenAI for investigative journalism. This is technology that can comb through complex datasets and help spot trends. What you are about to read about is one step further: agentic behaviour.

The second is chat products. I’ve written already about how chat is a trend to watch: An increasing number of news brands are building chat products using GenAI so their readers can quickly access exactly the information they are looking for (because, let’s face it, search functionality on most news Web sites is not very good). It is part of a broader trend towards personalisation and handing control to the user of finding relevant content in a format that works best for them.

Read on — and let me know what you’re working on or thinking about. I’d love to hear from you.

Sonali

GenAI for investigations

I regularly get questions from INMA members on the use of agents, so I was intrigued to come across a fascinating report that asks the question: “ChatGPT hasn’t quite hit the mark as an investigative reporting assistant — could an agentic AI workflow offer a better solution?”

The researchers developed a prototype system that, when provided with a dataset and a description of its contents, generates a “tip sheet” — a list of newsworthy observations that may inspire further journalistic explorations of datasets.

This system employs three AI agents, emulating the roles of a data analyst, an investigative reporter, and a data editor. “Just as human journalism benefits from collaboration, AI could also advance through teamwork,” wrote the authors of the report, researchers Joris Veerbeek and Nick Diakopoulos.

Overall, these three agents collaborate through four stages:

1. Question generation: First, a dataset and its description are provided to the reporter agent, which is tasked with brainstorming a set of questions (with the number adjustable) that could be answered using the data.

2. Analytical planning: For each question, the analyst drafts an analytical plan detailing how the dataset can be used to answer the question. The editor provides feedback on the plan and the analyst redrafts.

3. Execution and interpretation: Each analytical plan is executed and interpreted by the analyst. The editor and reporter provide feedback, which the analyst incorporates, and the reporter then summarises the final results in bullet points.

4. Compilation and presentation: All bullet points from the previous step are compiled, and a subset of the most significant findings is presented to the user in the tip sheet.

Also interesting is how they work together: Throughout these stages, the agents don’t just passively use each other’s outputs as inputs but actively have to incorporate each other’s feedback, particularly during the analysis phase. 

For example, after the analyst completes its work in the third step, the reporter steps in to assess these findings. The reporter is then prompted to choose between three choices:

1. Give a green light for “publication,” which signals the insight should be bulletproofed and potentially shared with the journalist supervising the agents.

2. Suggest further analysis to try to develop other angles.

3. Decide the findings aren’t newsworthy enough to pursue.

The team of agents was tested on five actual complex investigative data-journalism projects that had been nominated for awards.

“The results show that the process was surfacing leads with news potential which weren’t included in the original reporting,” the report said. “This means there is potential to inform avenues of investigation for new coverage.”

What is still not clear is how exactly the prompts, feedback loops, and knowledge bases contributed to the final outcome. 

“The system we’ve developed shows a lot of promise — it’s a tool that can help uncover valuable leads and provide new angles on complex stories,” the authors wrote. “But it’s also just that: a tool. The insights generated by these agents are a starting point, but the real work of journalism, the craft of telling a story that matters, remains firmly in human hands.”

Dates for the calendar: October 21-25

Join us for our Los Angeles Tech Innovation study tour! We will delve into best practices in audio and video content, new technologies, and different commerce models driving the creator economy to learn valuable lessons that can be applied to the news industry. I hope to see you there.

Three new chat products to watch

Here are three chat products that caught our eye in recent weeks:

Joannabot: This bot draws on the expertise of The Wall Street Journal’s tech columnist, Joanna Stern, and answers questions about the latest iPhone. After a week, the Journal ran a candid piece (all credit to them for their openness!) by Stern on how the bot was performing. 

Noteworthy: 

• They noticed the cost: “I could buy every reader of this column a granola bar for the cost of this project’s Google cloud bill,” Stern noted.

• The bot could be tricked into making inflammatory statements.

• It did hallucinate.

And yet, there were positives as well. “Joannabot could deliver personalised answers far faster than I ever could,” Stern wrote. “Plus, with no coding knowledge, I was able to improve its performance.”

Kamalabot: The San Francisco Chronicle built a chat product that answers its readers’ questions about U.S. presidential candidate Kamala Harris, drawing on 3,500 articles over three decades of the newspaper’s coverage of her since her career began in the region.

“The idea to build an archive-discovery tool focused on Harris sprung from an impromptu Sunday team call following the news that President Joe Biden was pulling out of the race for the presidency,” said Tim O’Rourke, vice president of content strategy. “What followed was weeks of collaboration, innovation, cautious decision-making, and buy-in from all levels of our organisation.”

Every question submitted goes through three checks before an answer is generated, the Chronicle tells readers. First, OpenAI’s moderation filter ensures the query does not contain offensive content, such as self-harm, hate speech, or violence. Then, the bot verifies whether the question pertains specifically to Harris. And finally, if the question is about where or how to vote, the reader is redirected elsewhere. 

“If no relevant information is found, the tool will suggest related articles instead of attempting to generate a response,” the Chronicle said. “To continually improve the service, we save questions submitted by readers. These questions help us update the FAQ section regularly and guide future reporting efforts.”

Leckerschmecker bot: Germany’s Funke Mediengruppe is weeks away from the launch of a new chat product that helps its readers find recipes on its Leckerschmecker food site. The reader can specify ingredients they would like to use and the bot will provide them with recipes. The reader can then ask for modifications, such as making a recipe vegetarian or requesting a substitute for an ingredient suggested. The bot will prompt them as well to seek further information.

The aim is to generate more pageviews per user and strengthen engagement with Leckerschmecker’s content since most readers come to the site, find a recipe, and leave.

“We think this will really drive engagement. Either they are happy with the recipe we give them — or they get another recipe from us that might be even more interesting for them,” said Paul Elvers, head of AI at Funke.

The bot will offer the reader many relevant recipe options for browsing as well as customisation options and will appear on both the Leckerschmecker homepage as well as on recipe article pages. 

The chat product will be monetised by advertisements on the article page. Funke is also considering integrating advertising into the bot itself, said Elvers.

Also on the road map: the ability for a user to upload a photograph of what is in their fridge and to ask what they should cook.

Worthwhile links

• GenAI tools: Which ones are good for what? Edelman put together a handy list. It is aimed at marketing communications professionals, but many of them are useful for news brands as well.
• GenAI adoption: Many INMA members struggle with getting their journalists on board with AI tools. Here is how Sweden's Omni does it.
• GenAI and decisions: Where to start? This interesting blog post outlines a smart way to approach GenAI use cases.
• GenAI and transparency: Humans don’t trust headlines labelled as AI-generated.
• GenAI and chat: Interesting read from Forrester on mistakes we make when designing chatbots.
• GenAI and monetisation: Google will place ads against its AI Overviews. (We knew this was coming.)
• GenAI and productivity: Three scenarios for the future from MIT’s Daron Acemoglu.
• GenAI and hallucinations: Can mathematics solve the problem?
• GenAI and electricity: Tech giants try to solve the problem of insatiable demand for power. Some of the numbers in this article are staggering.
• GenAI and small, open models: Draw some inspiration from scientists.
• GenAI and death: Thought working from home was a stretch? Now, you can work from the grave.

About this newsletter

Today’s newsletter is written by Sonali Verma, based in Toronto, and lead for the INMA Generative AI Initiative. Sonali will share research, case studies, and thought leadership on the topic of generative AI and how it relates to all areas of news media.

This newsletter is a public face of the Generative AI Initiative by INMA, outlined here. E-mail Sonali at sonali.verma@inma.org or connect with her on INMA’s Slack channel with thoughts, suggestions, and questions.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .